JP2001154685A - Device and method for voice recognition and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve entertaining capability of a robot by realizing a voice recognition process that changes in accordance with the growing state of the robot. SOLUTION: A growing score setting section 46 sets a growing score for each word registered in the word dictionary of a dictionary storage section 44 based on the growing state of the robot. A matching section 42 conducts voice recognition with respect to the words registered in the dictionary while considering the growing scores set to the words.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a voice recognition device, a voice recognition method, and a recording medium, and more particularly to a voice recognition device, a voice recognition method, and a recording medium suitable for a robot having a voice recognition function.

[0002]

2. Description of the Related Art For example, there is known a speech recognition technique in which an acoustic score and a linguistic score are obtained based on an acoustic model, a dictionary, a grammar, and the like, and a word sequence is determined as a speech recognition result of an input speech. In addition, the control method for controlling the operation of the robot in response to a certain command signal is called FA (Factor
y Automation). Further, a system that controls the operation of the robot based on the result of the voice recognition by combining the voice recognition technology and the robot control technology has been realized. As such a system, for example, there is a system in which sorting of packages is changed using voice recognition in sorting of packages. In addition, in addition to the robot, control of a computer (command input or the like) or the like using voice recognition has been widely performed in recent years.

[0003] In addition to the above-mentioned industrial robots, a robot for entertainment (including a stuffed animal) as a pseudo pet also recognizes a user's utterance as a speech and outputs the speech recognition result. On the basis of,
Products that perform various actions are being commercialized.

[0004]

In particular, it is desired for a robot for entertainment to allow the user to enjoy the same feeling as a real pet growing. That is, for example, at first, the utterance of the user cannot be understood (speech recognition), but as the user grows up, the utterance of the user can be gradually understood.
The entertainment of the robot can be significantly improved.

The present invention has been made in view of such a situation, and it is an object of the present invention to perform a voice recognition process that changes in accordance with the state of a robot, such as growth.

[0006]

According to the present invention, there is provided a voice recognition apparatus comprising: voice recognition means for recognizing voice; and control means for controlling the voice recognition means based on a state of the robot.

The control means can control the voice recognition means based on the growth, emotion, or state of the instinct of the robot.

The control means can change the speech recognition accuracy of the speech recognition means based on the state of the robot.

The speech recognition means includes dictionary storage means for storing a dictionary in which vocabularies to be subjected to speech recognition are registered.
The control means can control the voice recognition means so that the vocabulary registered in the dictionary is subjected to voice recognition with a weight corresponding to the state of the robot.

The voice recognition means includes dictionary storage means for storing a plurality of dictionaries in which vocabularies to be subjected to voice recognition are separately registered. The voice recognition means can be controlled so that voice recognition is performed with weights corresponding to the states.

Further, the speech recognition means is provided with a dictionary storage means for storing a dictionary in which the vocabulary to be subjected to speech recognition is registered in association with another vocabulary, and the control means is provided for controlling the state of the robot. Based on the dictionary, the speech recognition unit can be controlled so that another vocabulary associated with the vocabulary as the speech recognition result is output as the final speech recognition result.

In the dictionary, vocabulary to be subjected to speech recognition is
The vocabulary can be registered in association with another vocabulary acoustically similar or semantically related to the vocabulary.

[0013] The speech recognition means is provided with a dictionary storage means for storing a dictionary in which vocabularies to be subjected to speech recognition are registered.
The control means can control the number of words that can be registered in the dictionary based on the state of the robot.

[0014] The robot can take a predetermined action based on the voice recognition result output by the voice recognition means.

The voice recognition method according to the present invention includes a voice recognition step for recognizing voice, and a control step for controlling the voice recognition step based on a state of the robot.

[0016] The recording medium of the present invention is characterized in that a program comprising a voice recognition step for recognizing voice and a control step for controlling the voice recognition step based on the state of the robot is recorded.

In the voice recognition device, the voice recognition method, and the recording medium according to the present invention, voice recognition processing is controlled based on the state of the robot.

[0018]

FIG. 1 shows an example of the appearance of a robot according to an embodiment of the present invention, and FIG.
An example of the electrical configuration is shown.

In the present embodiment, the robot has a dog shape, a head unit 3 is disposed at the front end of the body unit 2, and a head unit 3 is provided at each of the front, rear, left and right corners. Are the thigh units 4A, 4B, 4C,
A leg unit 6A, 6B, 6C, 6D composed of 4D and shin unit 5A, 5B, 5C, 5D is attached. Further, a tail unit 1 is attached to a rear end of the body unit 2.

The tail unit 1 and the torso unit 2, the head unit 3 and the torso unit 2, the thigh units 4A to 4D and the torso unit 2, and the thigh units 4A to 4D and the shin unit 5A respectively. Through 5D
Motors 7 ₁ , 7 ₂ ,..., 7 _N as actuators are provided in the respective joint mechanisms connecting the respective motors, and by driving the corresponding motors 7 _{1 to} 7 _N , Tail unit 1 and head unit 3
Is freely rotated in directions around the three axes x, y, and z axes, and the thigh units 4A to 4D are
Axis and two axes of the y-axis can be freely rotated in the directions thereof, and the shin unit 5A to 5D can be rotated in the direction of one of the x-axes. The robot is capable of performing various actions.

The head unit 3 has a camera (CCD (Cha)
rge Coupled Device (camera) 8, a microphone (microphone) 9, and a pressure sensor 10 are respectively provided at predetermined positions, and the body unit 2 includes a control unit 1
1 is built in. The camera 8 captures an image of the surrounding situation including the user, and the microphone 9 collects surrounding sounds including the user's voice. Further, the pressure sensor 10 detects a pressure applied by a user or the like. Then, an image of the surrounding situation captured by the camera 8,
Ambient sound collected by the microphone 9, pressure sensor 10
Are supplied to the control unit 11 as image data, audio data, and pressure detection data, respectively.

Each motor 7 _{1 to} 7 _N corresponding to each joint mechanism
For are rotary encoders 12 ₁ to 12 _N are provided corresponding to each of the respective rotary encoders 12 ₁ to 12 _N, the rotation angle of the rotating shaft of the corresponding motor 7 ₁ to 7 _N is detected. Rotary encoder 12
_The rotation angles detected at _{1 to} 12 _N are given to the control unit 11 as angle detection data.

The control unit 11, the image data from the camera 8, based audio data from the microphone 9, a pressure detection data from the pressure sensor 10, and the angle detection data from the rotary encoder 12 ₁ to 12 _N respectively, around as well as determine the conditions and their orientation, etc. in advance based on the installed control program determines the subsequent actions, based on the determination result, so as to drive the required motor 7 ₁ to 7 _N I have.

Thus, the robot can move the tail unit 1
Alternatively, the user moves the head unit 2 and each of the leg units 6A to 6D to a desired state and acts autonomously.

FIG. 3 shows an example of the configuration of the control unit 11 shown in FIG.

The control unit 11 has a CPU (Central Process
ng Unit) 20, program memory 21, RAM (Rando
m Access Memory) 22, nonvolatile memory 23, interface circuit (I / F) 24, and motor driver 2
5 are connected via a bus 26.

CPU (Central Processing Unit) 20
Controls the behavior of the robot by executing a control program stored in the program memory 21. The program memory 21 is, for example, an EEPROM (E
lectrically Erasable Programmable Read Only Memor
y) and the like, and stores a control program executed by the CPU 20 and necessary data. RAM 22 is a CPU
20 temporarily stores necessary data and the like. The nonvolatile memory 23 stores an emotion /
Data such as an instinct model, a behavior model, a growth model, a word dictionary, a phonological model, a grammar rule, and the like, which need to be retained even after the power is turned off, are stored. Interface circuit 24, a camera 8 and a microphone 9, the pressure sensor 10, rotary encoder 12 ₁ to 1
The data supplied from 2 _N is received and supplied to the CPU 20. The motor driver 25 is a CPU
Under the control of 20, a drive signal for driving the motor 7 ₁ to 7 _N (drive), and supplies it to the motor 7 ₁ to 7 _N.

When the control program stored in the program memory 21 is executed by the CPU 20, the control unit 11 functionally controls the behavior of the robot as shown in FIG.

That is, FIG. 4 shows an example of a functional configuration of the control unit 11.

The sensor input processing unit 30 recognizes a stimulus given from the outside and an external state based on the outputs of the microphone 9, the camera 8, and the pressure sensor.
The information is supplied to the emotion / instinct model unit 32 and the action determination unit 33. The sensor input processing unit 30
Performs processing using the output of the growth model unit 31 as necessary. This will be described later.

The growth model unit 31 manages a growth model representing the growth state of the robot. The growth model unit 31 changes the growth state of the robot based on the output of the sensor input processing unit 30 and the passage of time. Is supplied to the sensor input processing unit 30 and the behavior determination unit 33.

The emotion / instinct model unit 32 manages an emotion model and an instinct model that express the emotion of the robot and the state of the instinct, respectively.
By changing the parameters defining the emotion model and the instinct model based on the output of the action determination unit 33 and the passage of time, the state of the emotion and the instinct of the robot is changed. The state of the emotion and the instinct after the change are supplied to the action determining unit 33.

The action determining section 33 includes an action model storage section 33
A, the memory contents, the output of the sensor input processing unit 30, the growth model managed by the growth model unit 31, the emotion model and the instinct model managed by the emotion / instinct model unit 32, and the time. Based on the progress, an action to be performed by the robot is determined, and the information (hereinafter, appropriately referred to as action information) is supplied to the posture transition unit 34.

The posture transition section 34 calculates the control data such as rotation angle or rotational speed of the motor 7 ₁ to 7 _N for causing the robot actions corresponding to the action information supplied from the action decision unit 33, Output to the motor control unit 35.

The motor controller 35 drives the motors 7 _{1 to} 7 _N to rotate in accordance with the control data from the attitude transition unit 34.

In the control unit 11 configured as described above,
In the input sensor processing unit 30, image data supplied from the camera 8, audio data supplied from the microphone 9,
Based on the pressure detection data output from the pressure sensor 10, a specific external state, a specific action from the user,
An instruction or the like from the user is recognized, and the recognition result is output to the growth model unit 31, the emotion / instinct model unit 32, and the action determination unit 33.

That is, the sensor input processing unit 30
Performs image recognition on the basis of the image data output by the device, and recognizes, for example, "there is a ball" or "there is a wall", and outputs the recognition result to the growth model unit 31, the emotion / instinct model unit. 32 and an action determining unit 33. The sensor input processing unit 30 performs voice recognition based on the voice data output from the microphone 9, and outputs the voice recognition result to the growth model unit 31 and the emotion / instinct model unit 3.
2, and the action determination unit 33 is supplied. Further, the sensor input processing unit 30 processes the pressure detection data given from the pressure sensor 10 and, for example, when it detects a pressure that is equal to or more than a predetermined threshold value and for a short time, “hits”. When the pressure is detected below a predetermined threshold value and for a long time, it is recognized as “patched (praised)”, and the recognition result is used as the growth model unit 3
1, supply to the emotion / instinct model unit 32 and the action determination unit 33.

The growth model unit 31 manages a growth model for defining the growth state of the robot. The growth model is composed of, for example, an automaton as shown in FIG. In this automaton, the growth state is represented by nodes (states) NODE _{0 to} NODE _G , and the growth,
That transition growth state, the node NODE _g corresponding to growth state, the node NO corresponding to the next growth state
It is represented by an arc ARC _{g + 1} representing a transition to DE _{g + 1} (g = 0, 1,..., G-1).

In this embodiment, the state of growth
Is shown in FIG. 5A (the same applies to FIG. 5B described later).
Transitions from the left node to the right node
It has become. Therefore, in FIG.
For example, the leftmost node NODE ₀Is just born
Indicates the state of "newborn" and the second node NOD from the left
E₁Represents the state of "infant", and the third node from the left
NODE_TwoIndicates the state of "child". The same applies to the following
The more nodes to the right, the more grown
Also the right node NODE_GIndicates an "elderly" state.

An arc ARC _{g + 1} representing a transition from a certain node NODE _g to the right node NODE _{g + 1} is set with a condition (input) P _{tg + 1} for _{causing the} transition. The transition (growth) of a node is determined based on this condition. That is, in the arc ARC _{g + 1} ,
The condition P _{tg + 1} regarding the output from the camera 8, the microphone 9, the pressure sensor 10, the passage of time, and the like required to cause the transition are set, and the condition P _{tg + 1} is satisfied. in the case, the node nODE _g from the right adjacent node NO
The transition to DE _{g + 1} occurs and the robot grows.

Note that, in the growth model shown in FIG. 5A, since there is only one fixed path that transitions from the left node to the right node, the growth process is performed on that one path. According to the growth model, a plurality of growth processes exist, for example, as shown in FIG.
It is also possible to use the one shown in FIG.

That is, in FIG. 5B, the growth model has a tree structure. For example, from the node NODE ₀ corresponding to “newborn”, the nodes NODE _1-0 corresponding to “infant” or It may transition to any of NODE _1-1 . Two nodes NODE _1-0 corresponding to "infants"
And NODE _1-1 represent, for example, in human terms, infants of different personalities, nodes NODE _1-0 or N
Which one of the ADEs _1-1 to transition to is determined by the camera 8
Or based on the output from the microphone 9, the pressure sensor 10, the passage of time, and the like. Therefore, in this case, the growth process of the robot can be made different depending on the received stimulus or the like.

As a growth model, in addition to the tree structure,
A structure that can be expressed by an automaton such as a graph structure can be adopted.

The growth model unit 31 manages the growth state of the robot based on the above growth model, and displays the current growth state on the sensor input processing unit 30 and the action determination unit 3.
Output to 3.

On the other hand, the emotion / instinct model section 32 is used to control the robot, for example, to be "happy", "sad", "angry", etc.
And emotion models expressing so-called emotional and emotional states, such as “appetite”, “sleep desire”, and “exercise desire”
And an instinct model expressing the state of the instinct.

That is, the emotion model and the instinct model indicate the emotion and the state of the instinct of the robot, for example, from 0 to 1 respectively.
The emotion / instinct model unit 32 changes the values of the emotion model and the instinct model based on the output of the sensor input processing unit 30, the output of the action determination unit 33, and the passage of time. Let it. Therefore, here, the state of the robot's emotions and instinct changes based on the output of the action determining unit 33 and the passage of time, in addition to the external recognition result output by the sensor input processing unit 30. Then, the emotion / instinct model unit 32 stores the values of the emotion model and the instinct model after the change (the emotions of the robot and the state of the instinct) into the action determination unit 3
Supply 3

Note that the emotion model and the instinct model can also be constituted by, for example, an automaton or the like.

Here, the state of the emotion and the instinct of the robot changes as follows, for example, based on the output of the action determining unit 33.

That is, the action determination section 33 indicates the current action or the action taken in the past (for example, "turned away") of the robot to the emotion / instinct model section 32. Behavior information is provided.

Now, suppose that the robot receives a certain stimulus from the user when the feeling of "anger" is high, and the robot takes an action "turns in the direction". In this case, the behavior information “turned away” is supplied from the behavior determination unit 33 to the emotion / instinct model unit 32.

In general, when an angry person takes an action that expresses dissatisfaction such as turning to the side, anger may be somewhat calmed down by the action. Therefore, when the behavior information “turned away” is supplied from the behavior determination unit 33, the emotion / instinct model unit 32 changes the value of the emotion model representing “anger” to a small value (“anger”). To reduce the degree of

The action determining unit 33 includes the sensor input processing unit 30
Based on the external recognition result output by the user, the growth model in the growth model unit 31, the emotion model and the instinct model in the emotion / instinct model unit 32, the storage contents of the action model storage unit 33A, the elapsed time, and the like. And determine
The behavior information representing the behavior is supplied to the emotion / instinct model unit 32 and the posture transition unit 34.

Here, the behavior model storage unit 33A stores a behavior model that defines the behavior of the robot, for example, as shown in FIG.

That is, in the present embodiment, the behavior model is
For example, it is composed of a stochastic automaton. In this stochastic automaton, an action is represented by nodes (states) NODE _{0 to} NODE _M , and a transition of an action is performed from a node NODE _m0 corresponding to a certain action to another action (the same action may be performed). It is represented by an arc ARC _m1 representing a transition to the corresponding node NODE _m1 (m0, m1 = 0,
1,..., M).

In FIG. 6, in order to avoid complication, the stochastic automaton composed of M + 1 nodes is transferred from node NODE ₀ to other nodes (including itself) NODE _{0 to} NODE _M. Arc ARC ₀
ARC _M only are shown.

From one node NODE _m0 to another node N
An arc ARC _m1 representing a transition to ODE _m1 is set with a transition probability P _{tm1 at} which the transition occurs, and each node NODE _m (m = 0, 1,..., M) has An occurrence probability P _{nm at} which an action corresponding to the node NODE _m occurs is set. In the action model, what action to take after a certain action is determined from the transition probability set for the arc and the occurrence probability set for the node.

The action determining unit 33 sets a transition from a node corresponding to the current action (hereinafter, appropriately referred to as a current action) to an arc extending from the node corresponding to the current action in the action model. In addition to the transition probabilities set and the occurrence probabilities set in the transition destination node indicated by the arc, the growth model in the growth model unit 31, the emotion model and the instinct model in the emotion / instinct model unit 32, and the time course , Sensor input processing unit 30
Is determined based on the external recognition result output by the node, and an action corresponding to the node after the transition (hereinafter, appropriately referred to as a transition action)
Is supplied to the emotion / instinct model unit 32 and the posture transition unit 34.

Accordingly, even if the external recognition results output by the sensor input processing unit 30 are the same, for example, the behavior determining unit 33 may take different actions depending on the growth state represented by the growth model. It is determined.

Here, the behavior determining unit 33 changes the occurrence probability set for each node of the behavior model based on the growth model.

More specifically, for example, ignoring the transition probability set for each arc in the action model and considering only the occurrence probability set for each node, for example, the occurrence probability set for each node is considered. The sum is 1. For example, assuming that there are four behaviors, “sleep”, “rise”, “walk”, and “run”, in the behavior model, if the occurrence probability of a certain behavior is represented by P (behavior), P (“Sleep”) + P (“stand up”) + P (“walk”) + P (“run”) is 1.

Depending on the growth of the robot, the sum of the occurrence probabilities set for each node remains the same, but the action determination unit 33 determines that the occurrence probability of the action corresponding to each node is Therefore it is changed.

That is, for example, in the growth model, when the growth state represents “newborn”, the behavior determining unit 33 sets the occurrence probability of the behavior “sleep” to 1, and sets the other behaviors “rise” and “ The occurrence probabilities of “walk” and “run” are set to 0. In this case, the robot does not take any action other than “sleep”.

After that, when the growth state becomes “infant” in the growth model based on the lapse of time or the like, the behavior determining unit 33 reduces the occurrence probability of the behavior “sleep” and reduces the probability of the behavior “rise”. Increase the probability of occurrence. Furthermore, as the robot grows, the behavior determining unit 33 gradually increases the occurrence probability of each of the behaviors “walk” and “run”. This allows the robot to initially only "sleep", but "grow up" as it grows up, and "walk"
You will take actions such as running. That is, possible behavior patterns increase.

In the action determining unit 33, as described above,
Even if the external recognition results output by the sensor input processing unit 30 are the same, it is determined that different actions are taken depending on the growth state, and different actions are taken depending on the values of the emotion model and the instinct model. Is determined.

Specifically, for example, the sensor input processing unit 3
In the case where the output of 0 indicates that "the palm is put in front of the eyes", the emotion model of "anger" indicates "not angry" and the instinct of "appetite" When the model indicates “not hungry”, the action determining unit 33 determines to perform “hand” as a transition action in response to the palm being presented in front of the eyes. I do.

For example, as in the case described above, the output of the sensor input processing unit 30 indicates that "the palm is put in front of the eyes" and the emotion model of "anger" is "Not angry" means "appetite"
If the instinct model indicates that the user is hungry, the action determining unit 33 determines to perform an action such as “licking the palms” as the transition action.

Further, the output of the sensor input processing unit 30 is
It indicates that "the palm is in front of you", but if the emotion model of "anger" indicates "angry", the value of the instinct model of "appetite" Regardless, the action determining unit 33 determines that an action such as “turn to the side” is performed as the transition action.

Note that the action determining unit 33 outputs a predetermined trigger (t
rigger), transition the nodes of the behavior model and determine the transition behavior. That is, for example, when a predetermined time has elapsed since the robot started the current action, or when a specific recognition result such as a voice recognition result is output from the sensor input processing unit 30, When the growth model in the model unit 31 represents a certain growth state, the transition behavior is determined when the value of the emotion model or the instinct model in the emotion / instinct model unit 32 reaches a predetermined threshold.

The action determining unit 33 recognizes the current posture of the robot on the basis of the outputs of the rotary encoders 12 _{1 to} 12 _N , and performs a transition action in a natural manner from the posture. The behavior information is transferred to the posture transition unit 34
Output.

The posture transition unit 34 generates posture transition information for transitioning from the current posture to the next posture based on the behavior information supplied from the behavior determining unit 33, and outputs it to the motor control unit 35. That is, the posture transition unit 34 determines the current posture via the action determination unit 33 and the rotary encoder 12.
The rotation angles and rotation speeds of the motors 7 _{1 to} 7 _N for recognizing based on the outputs of _{1 to} 12 _N and causing the robot to take an action (transition action) corresponding to the action information from the action determining unit 33 are described. The calculated value is output to the motor control unit 35 as attitude transition information.

The motor control unit 35 generates a drive signal for driving the motors 7 _{1 to} 7 _N according to the posture transition information from the posture transition unit 34 and supplies the drive signals to the motors 7 _{1 to} 7 _N. Thus, the motors 7 _{1 to} 7 _N are driven to rotate, and the robot performs a transition action.

Next, FIG. 7 shows the sensor input processing unit 3 of FIG.
0 shows a functional configuration example of a portion that performs voice recognition using voice data from the microphone 9 (hereinafter, appropriately referred to as a voice recognition device).

In this speech recognition apparatus, the speech input to the microphone 9 is, for example, a continuous distribution HMM (Hidden Matrix).
(rkov Model) method, and the speech recognition result is output.

That is, audio data from the microphone 9 is supplied to the characteristic parameter extraction unit 41. The characteristic parameter extraction unit 41 converts the audio data input thereto into appropriate frames. , For example, M
FCC (Mel Frequency Cepstrum Coefficient) analysis is performed, and the analysis result is used as a feature parameter (feature vector).
Is output to the matching unit 42. The feature parameter extraction unit 41 can also extract, for example, a linear prediction coefficient, a cepstrum coefficient, a line spectrum pair, power (output of a filter bank) for each predetermined frequency band, and the like as feature parameters. .

The matching unit 42 uses the feature parameters from the feature parameter extraction unit 41 to input to the microphone 9 while referring to the acoustic model storage unit 43, the dictionary storage unit 44, and the grammar storage unit 45 as necessary. The input speech is subjected to speech recognition based on the continuous distribution HMM method.

That is, the acoustic model storage unit 43 stores acoustic models representing acoustic features such as individual phonemes and syllables in the language of the speech to be recognized. Here, since speech recognition is performed based on the continuous distribution HMM method, HMM (Hidden Markov Model) is used as an acoustic model. The dictionary storage unit 44 stores, for each word to be recognized,
A word dictionary in which information (phonological information) related to the pronunciation is described is stored. The grammar storage unit 45 includes the dictionary storage unit 35
Grammar rules that describe how the words registered in the word dictionary are linked (connected). Here, the grammar rules include, for example, context-free grammar (CFG) and statistical word chain probability (N-gram).
Rules based on such as can be used.

The matching section 42 refers to the word dictionary in the dictionary storage section 44 and connects the acoustic models stored in the acoustic model storage section 43 to form a word acoustic model (word model). . Further, the matching unit 42 connects some word models by referring to the grammar rules stored in the grammar storage unit 45, and uses the word models connected in this manner, based on the feature parameters, The voice input to the microphone 9 is recognized by the continuous distribution HMM method. That is, the matching unit 42
Detects a sequence of a word model having the highest score (likelihood) at which a time-series feature parameter output from the feature parameter extraction unit 41 is observed, and recognizes a word string corresponding to the sequence of the word model as speech recognition. Output as result.

More specifically, the matching unit 42 accumulates the appearance probabilities of the characteristic parameters input in time series with respect to the word string corresponding to the connected word model, and uses the accumulated value as a score. The highest word string is output as a speech recognition result. The speech recognition result by the matching unit 42 is output as an output of the sensor input processing unit 30 to the growth model unit 31, the emotion / instinct model unit 32, and the action determination unit 33, as described above.

Here, if a score is calculated for a word string that can be obtained by all the words registered in the word dictionary in the dictionary storage unit 44, the amount of calculation becomes enormous. Whether or not to calculate a score for is limited by the grammar rules stored in the grammar storage unit 45. In the matching unit 42,
In some cases, weights may be assigned to increase the score of a word string having a high frequency of appearance.

The score calculation in the matching unit 42 is as follows.
For example, an acoustic score given by an acoustic model stored in the acoustic model storage unit 43 (hereinafter, appropriately referred to as an acoustic score) and a linguistic score given by a grammar rule stored in the grammar storage unit 45 (hereinafter, referred to as an acoustic score) , A language score, as appropriate).

That is, assuming that the k-th word in a word string composed of a certain N words is W _k , the acoustic score of the word W _k is A (W _k ), and the language score is L (W _k ). When represented, the score S of the word string is calculated, for example, according to the following equation.

S = Σ (A (W_k) + C_k× L (W_k)) (1) where Σ is the summation of changing k from 1 to N
It represents that. Also, C _kIs the word W_kLanguage score L
(W_k).

The dictionary storage unit 4 included in the speech recognition device
The word dictionary stored in No. 4 defines the vocabulary that can be handled by the robot, and is very important in that respect.
That is, as the number of words registered in the word dictionary increases, the robot can deal with various utterances.

However, the speech recognition device shown in FIG.
The function is realized by the U20 executing the program. In this case, the number of vocabularies in the word dictionary is limited by the processing capacity of the CPU 20, the size of the free area of the RAM 22 available to the CPU 20, and the like. receive.

Further, even if, for example, the same vocabulary is uttered by the same person, the voice waveform hardly matches completely, so that the matching process (score calculation) in the matching unit 42 is ambiguous. In some cases, erroneous recognition in which an unuttered word string is regarded as a speech recognition result may occur due to the processing of an input having a characteristic, but such erroneous recognition is generally caused by the number of vocabularies registered in the word dictionary. Increases, the number of similar words also increases.

Therefore, the vocabulary registered in the word dictionary is CP
In addition to the processing capacity of the U20 and the capacity of the RAM 22, the required recognition accuracy is limited.

The growth score setting section 46 controls the growth model section 3
The speech recognition processing by the speech recognition device is controlled based on the growth model in 1. That is, in the embodiment of FIG. 7, the growth score setting unit 46 sets the
The word dictionary in the dictionary storage unit 44 is operated, thereby controlling the speech recognition processing.

Specifically, for example, the growth score setting unit 4
6 calculates, based on the growth model, a growth score indicating the degree of understanding of each word registered in the word dictionary in the growth state of the robot. For example,
Now, assuming that the probability that a certain word W can be understood in the growth state of the robot is P (W), the growth score setting unit 46 sets the growth score G (W) for the word W as
For example, it is calculated according to the following equation.

G (W) = − log (P (W)) (2)

Further, the growth score setting section 46 associates the growth score with a corresponding word in the word dictionary in the dictionary storage section 44. Therefore, the word dictionary generally stores information for identifying a word such as a word notation and phonological information of the word. The word dictionary in the dictionary storage unit 44 is, for example, As shown in FIG. 8, the growth score for the word is registered in addition to the word notation and phonological information. Here, in FIG. 8, the words “run”, “get up”, “walk”, “hand”, “sit”,.
The growth scores of 0, 40, 50,... Are registered.

In the word dictionary in the dictionary storage unit 44, FIG.
As shown in (1), when the growth score is registered, the matching unit 42 obtains the score for the word string in consideration of not only the phonological score and the language score but also the growth score. That is, in the matching unit 42, the expression (1)
, The score S is calculated according to, for example, Expression (3).

S = Σ (A (W _k ) + C _k × L (W _k ) + D _k × G (W _k )) (3) where G _k (W _k ) is the k-th word W _represents a growth score for _k , and D _k represents a weight applied to the growth score G (W _k ).

As described above, when calculating the score S in the matching unit 42, the accuracy of speech recognition by the speech recognition device is changed according to the growth state of the robot by introducing a growth score.

That is, for example, when the growth state is “infant”, the growth score of the word “run” in the word dictionary of FIG. 8 is made smaller than the growth scores of the other words. , The speech recognition accuracy of the utterance “run” is lowered. As a result, for example, when "Run" is uttered, it is common to make the robot take the action of "Run" by accurately recognizing the utterance as a voice, but here, so to speak, intentionally Then, the utterance "run" is erroneously recognized, and the robot is prevented from taking the action of "running" (or is caused to take an action other than "running").

Then, as the robot grows, the growth score of the word “run” is gradually increased, and as a result, the utterance “run” is accurately recognized by speech, and the robot follows the utterance. , "Run".

Here, for example, in the case of a human being, the vocabulary that can be understood increases as the person grows up (or learns). Therefore, the speech recognition is performed based on the growth score as described above. Changing the accuracy according to the growth model will be consistent with actual human growth.

Next, referring to the flowchart of FIG.
The operation of the voice recognition device of FIG. 7 will be further described.

A voice is input to the microphone 9, and digital voice data corresponding to the voice is input to the sensor input processing unit 3.
When the audio data is supplied to 0, the audio data is supplied to the feature parameter extraction unit 41.

In the feature parameter extracting section 41, in step S 1, the voice data is subjected to acoustic analysis, whereby the feature parameters are extracted in time series and supplied to the matching section 42. In step S2, the matching unit 42 uses the feature parameters from the feature parameter extraction unit 41 to calculate (match) a score in consideration of the growth score according to Expression (3), and based on the score, 9 to determine the speech recognition result of the speech input. Then, the process proceeds to step S3, where the matching unit 42 outputs the speech recognition result to the growth model unit 31,
The output is sent to the emotion / instinct model unit 32 and the action determination unit 33, and the process ends.

In the above case, by changing (adapting) the growth score of a word registered in the word dictionary according to the growth state of the robot obtained from the growth model, the speech recognition accuracy of the speech recognition apparatus is changed. However, the speech recognition accuracy may be changed by changing, for example, an acoustic model in the acoustic model storage unit 43 or a parameter defining a grammar rule stored in the grammar storage unit 45 in accordance with a growth state of the robot. It is also possible to change it by doing.

In the speech recognition device, speech recognition is performed on words registered in the word dictionary.
By changing the words registered in the word dictionary, it is possible for the voice recognition device to change words to be subjected to voice recognition. That is, for example, in accordance with the growth state of the robot obtained from the growth model, words are sequentially registered in the word dictionary, or words registered in the word dictionary are deleted, and words to be subjected to speech recognition are changed. It is possible to do so.

Further, in the above case, the voice recognition accuracy of the voice recognition device is changed according to the growth state of the robot. However, the voice recognition accuracy is changed according to the emotion and the state of the instinct of the robot. It is also possible to make it. When the voice recognition accuracy is changed according to the state of the robot's emotion or instinct, the emotion / instinct model unit 32 sends the emotion model and instinct to the sensor input processing unit 30 as shown by the dotted line in FIG. To supply the model,
As shown by the dotted line in FIG.
6, the emotion score and the instinct score similar to the growth score are obtained from the emotion model and the instinct model and registered in the word dictionary. In this case, the matching unit 42 calculates the final score in consideration of the emotion score and the instinct score, so that the voice recognition accuracy can be changed according to the emotion and the state of the instinct of the robot. .

That is, when the emotion score and the instinct score are taken into consideration, when the robot is in a good mood, for example, the speech recognition accuracy of the words “hand” and “sit” is improved. "Hands" and "sitting" can be correctly recognized by speech, and corresponding actions can be taken.
Also, when the robot is in a bad mood, for example,
Decreases the voice recognition accuracy of the words "hand" and "sitting"
This makes it possible to misrecognize the utterance “hand” or “sitting”, cause an action unrelated to the utterance, and express the discomfort of the robot.

In addition to the human voice, the voice recognition device shown in FIG. 7 includes, for example, the sound of applause, the sound of hitting each part of the robot, the sound of musical instruments, the environmental sound generated around the robot, and the music ( Various sounds such as music) can be recognized as voice. That is, for example, when speech recognition is performed by the continuous distribution HMM method, the speech (sound) to be subjected to speech recognition may be modeled in the HMM. And
Even in the case of performing voice recognition on voices (sounds) other than human voices, the voice recognition accuracy can be changed according to the robot's growth state, emotional state, and instinct state.

Next, when performing speech recognition, all words to be subjected to speech recognition are registered in one word dictionary, and words to be subjected to speech recognition are divided into a plurality of groups ( However, the same word may belong to different groups), and for each word belonging to each group,
Sometimes a word dictionary is constructed. As described above, when a plurality of word dictionaries are prepared by dividing words to be subjected to speech recognition into a plurality of groups, speech recognition processing for a certain utterance is generally performed not on all of the plurality of word dictionaries but on the whole. Is performed using one or two or more word dictionaries.

That is, the number of words to be subjected to speech recognition for a certain utterance is limited by the required recognition accuracy in addition to the processing capacity of the CPU 20 and the capacity of the RAM 22 as described above.

Therefore, a plurality of word dictionaries are prepared,
In some cases, voice recognition of a certain utterance is performed by selecting a word dictionary to be used based on required recognition accuracy or the like.

Also, words to be subjected to voice recognition are grouped for each domain (theme) such as travel, baseball, news, etc., to form a plurality of word dictionaries.
Select a word dictionary to be used according to the situation where speech recognition is performed, and improve the speech recognition performance of proper nouns (for example, names of regions, names of baseball players, company names, etc.) included in a specific theme. In some cases, it can be improved.

Thus, a plurality of word dictionaries are constructed,
Selecting a word dictionary to be used and performing speech recognition is performed by using the method shown in FIG.
The present invention can also be applied to a voice recognition device.

That is, for example, the words to be subjected to speech recognition are now divided into Q groups, whereby the Q word dictionaries D ₁ , D ₂ ,..., D _Q corresponding to each group are obtained.
And stored in the dictionary storage unit 44.

Further, the growth score setting section 46 is associated with a growth score to be assigned to each of the _Q word dictionaries D _{1 to} D _Q according to the growth state of the robot. For example, as shown in FIG. A new dictionary growth score table is created and stored.

Here, in FIG. 10, for example, FIG.
A growth score for each of the word dictionaries D _{1 to} D _Q is set for the growth state of the robot represented by each node of the growth model shown in FIG.

According to the dictionary growth score table shown in FIG. 10, from the current robot growth state (node representing),
A growth score for each of the word dictionaries D _{1 to} D _Q is determined.

The growth score setting unit 46 registers the growth score determined based on the dictionary growth score table for each of the words in the word dictionaries D _{1 to} D _Q in the same manner as described above, and Then, a matching process is performed in consideration of the growth score.

If a certain word is registered in two or more word dictionaries, for example, the largest value among the growth scores for the two or more word dictionaries is registered in the word. You. In this case, for example, a word having a growth score of 0 is not targeted for speech recognition. That is, here, speech recognition is performed only for words registered in a word dictionary in which a growth score greater than 0 is set.

As described above, by setting a growth score for the word dictionary, the speech recognition accuracy can be changed according to the growth state of the robot.

[0117] That is, for example, now, in the dictionary growth score table of FIG. 10, when the growth status is "neonatal", together with the growth scores for the word dictionary D ₁ and 70, the growth score for the word dictionary D ₂ 30 and
Growth scores for other word dictionary D ₃ to D _Q keep zero. Further, when the growth state is “infant”, the growth scores for the word dictionaries D _{1 to} D ₃ are 40, respectively.
With a 10, 50, keep the 0 growth score for another word dictionary D ₄ to D _Q.

In this case, when the growth state is “newborn”, speech recognition is performed only on the words registered in the word dictionary D ₁ and the words registered in the word dictionary D ₂ . Therefore, if the words registered in the word dictionaries D ₁ and D ₂ are, for example, only three words “yes yes”, “papa”, and “mama”, only the utterance corresponding to these three words is accurate. Is recognized.

In this case, the growth score for the word dictionary D _{1 is set} to 70, and the growth score for the word dictionary D _{2 is set} to 30, so that the word dictionary D ₁
Towards growth score is larger than the growth score of the word dictionary D _2, from the words registered in the word dictionary D _2, found the following words registered in the word dictionary D _1, more accurately speech Be recognized.

[0120] Then, the robot grows, the growth state becomes "infant", another word dictionary D ₁ and D _2, even words registered in the word dictionary D ₃ speech recognition is performed as a target. Therefore, if the words registered in the word dictionary D ₃ are, for example, two words “hand” and “sitting”, the words “yes yes”, “papa”, and “mama” are three words. In addition, utterances corresponding to “hand” and “sitting” can be accurately recognized.

In this case, the growth state is “newborn”
From the word "infant", the word dictionary D ₁Growth against
A from 70 to 40, word dictionary D_TwoGrowth score for
Decreases from 30 to 10, respectively. Therefore, the word letter
Letter D₁And D_TwoThe recognition accuracy of words registered in
It decreases when the condition changes from "newborn" to "infant". Further
If the growth state is "infant", the word dictionary D_ThreeTo
Growth score is 50, and the word dictionary D₁And D_TwoTo
Greater than any of the growth scores. As a result,
When the long state becomes "infant", as described above, "Yes
I "," Daddy ", and" Mama "
Speech corresponding to "hand" or "sitting" can also be recognized.
Is possible, but "yes yes", "papa", and
The speech recognition accuracy of the utterance corresponding to "Mama" is
Deterioration of speech utterances for "sit" and "sit down"
I do. In other words, as they grow, robots say, "Yes
Used in early childhood such as "I", "Dad", and "Mama"
You will not be able to understand the word.

As described above, configuring a plurality of word dictionaries, selecting a word dictionary to be used, and performing speech recognition is as follows.
When there is a restriction on hardware or the like, it is convenient because speech recognition can be performed within the range of the restriction.

When a plurality of word dictionaries are prepared and speech recognition is performed as described above, a growth score is assigned to each word dictionary, and the emotion score and the instinct score described above are assigned. Is also possible.

When a tree model as shown in FIG. 5B is used instead of the growth model shown in FIG. 5A, the dictionary growth score table shown in FIG. It is possible to assign a growth score of the word dictionary to a growth state corresponding to each node in (B). in this case,
In FIG. 5B, for example, the growth state of “infant”
There are two states of nodes NODE _1-0 and NODE _1-1, even in growing conditions of the same "infant", the node NODE _1-0
And the node NODE
_The speech recognizable word and the speech recognition accuracy can be different between the case where the growth state corresponds to _1-1 .

That is, for example, when the growth state “newborn” corresponding to the node NODE ₀ grows to the growth state “infant” corresponding to the node NODE _1-0 , the voice is output when the growth state is “newborn”. In addition to the recognizable utterances “Yes Yes”, “Dad”, and “Mama”, utterances corresponding to “Hand” and “Sit” can be recognized by voice. On the other hand, when the child grows from the growth state “newborn” corresponding to the node NODE ₀ to the growth state “infant” corresponding to the node NODE _1-1 , voice recognition is possible when the growth state is “newborn”. The utterances "Yes Yes", "Dad", and "Mama" can be made speech-recognizable and only the speech corresponding to "hand" or "sitting" can be speech-recognized.

Next, in the speech recognition apparatus of FIG. 7, an accurate speech recognition result cannot be obtained for an utterance corresponding to a word (unknown word) not registered in the word dictionary in the dictionary storage unit 44.

Therefore, in order to enable accurate speech recognition of an unknown word, a new word can be registered in the word dictionary. Registration of words in the word dictionary is performed, for example, by providing a robot with an input interface, connecting a keyboard or the like, and operating the keyboard to give new word expressions and pronunciations (phonemes). It is possible to do.

As described above, the number of words registered in the word dictionary and subjected to voice recognition can be determined by the CPU 20 or the RAM 2.
2. In some cases, the number of words to be newly registered in the word dictionary is limited within a range that is not restricted by such problems. it can.

That is, the number of words newly registered in the word dictionary can be limited according to the growth state of the robot. Specifically, for example, when the growth state is "newborn", the number of words that can be newly registered is set to several words, and is increased to tens or hundreds of words as the robot grows. Can be. In this case, the words that the robot can understand will increase with growth.

Note that the number of words newly registered in the word dictionary can be changed based on the emotion score and the instinct score described above.

The sensor input processing unit 30 recognizes an image from the camera 8 in addition to the voice recognition. In this image recognition, in addition to a pattern such as a color and a face which can be recognized in advance, the image is recognized later. Thus, it is possible to newly register a color or a face pattern to be subjected to image recognition. Also in this case, the number of colors and face patterns that can be newly registered can be changed according to the growth state of the robot and the state of emotion or instinct.

Next, in the speech recognition apparatus of FIG. 7, as described above, the word string having the highest score is output as the speech recognition result. In general, it is desirable that the speech recognition result be accurate.

However, FIG.
In the voice recognition device of the above, it is possible to make the interaction with the robot fun by intentionally outputting an erroneous recognition result. That is, for example, in the case of a comic artist, the conversation partner views a word spoken by the speaker as an acoustically similar word, and produces a so-called “bokeh”. In some cases, so-called "tsukkomi" is performed, but in a voice recognition device, such "blurring" or "tsukkomi" can be performed by a robot by intentionally outputting an incorrect voice recognition result. Becomes

“Blur” can be realized, for example, as follows.

That is, in the word dictionary, as shown in FIG.
Words that are acoustically similar to the word but have a different meaning,
Alternatively, a word that is not acoustically similar to the word but is associated with the word (hereinafter referred to as replacement word information as appropriate)
Are registered in association with each other. Then, the matching unit 42
, Instead of words registered in the word dictionary obtained as a result of speech recognition, replacement word information associated with the words is output as a final speech recognition result.

In this case, since the action determining unit 33 determines the action to be performed next based on the replacement word information as the final speech recognition result, the robot does not correspond to the utterance content of the user. You will be performing a blurred action.

In addition, such a “bokeh” is prepared by preparing a plurality of word dictionaries as described above, for example.
By performing speech recognition using a certain word dictionary and using another word dictionary, other words that are acoustically similar to the words spoken by the user can be obtained as speech recognition results. This can also be realized.

Here, as described above, the voice recognition device
Intentionally outputting an erroneous speech recognition result is performed randomly and rarely using, for example, a state transition model having two states of “blurred” and “blurred”. It is desirable. Further, in this case, the way of state transition in the state transition model can be changed based on, for example, a growth model, an emotion model, or an instinct model.

Next, when the user makes an utterance that is “blurred”, the robot responds to the “blurred”
Performing "tsukkomi" actions, for example,
This is possible as in the case of realizing the above-described “blur”.

That is, it is difficult for the voice recognition device to recognize whether or not the user has made a "blurred" utterance (erroneous utterance). Therefore, as shown in FIG. Replacement word information is associated with the notation and phonological information of the word and registered in the word dictionary, and the matching unit 42 uses the word, not the word registered in the word dictionary, obtained as a speech recognition result, but the word. The associated replacement word information is output as the final speech recognition result.

When the user utters a “blurred” utterance, the replacement word information as the speech recognition result output by the speech recognition device should correspond to the utterance, and the user should originally utter the utterance. Since the action determining unit 33 determines the next action to be taken based on such replacement word information, the robot determines the action of "tsukkomi" in response to the user's "blurred" utterance. Will be done.

As described above, since the voice recognition processing is controlled based on the state of the robot, the voice recognition processing that changes in accordance with the state of the robot, such as growth, is performed. It is possible to provide a high robot and the like.

In this embodiment, the case where the present invention is applied to an entertainment robot (robot as a pseudo pet) has been described. However, the present invention is not limited to this. It can be widely applied to various robots such as robots.

[0144] Further, in this embodiment, has been to use a motor 7 ₁ to 7 _N as a driving means for causing an action to the robot, the present invention is not limited to this, short, with respect to the outside world Other actuators, speakers, buzzers, lighting devices, and the like can be widely used as the driving means as long as the action (operation) that brings about the action can be realized.

Furthermore, in the present embodiment, the above-described series of processing is performed by causing the CPU 20 to execute a program, but the series of processing may be performed by dedicated hardware. .

The program is stored in the program memory 21 (FIG. 3) in advance, and is stored on a floppy disk, CD-ROM (Compact Disc Read Only Memory), M
O (Magneto optical) disc, DVD (Digital Versatile)
Disc), a magnetic disk, a semiconductor memory, or another removable recording medium, which can be temporarily or permanently stored (recorded). Then, such a removable recording medium can be provided as so-called package software, and can be installed in the robot (program memory 21).

In addition to installing the program from a removable recording medium, the program can be downloaded from a download site.
It can be transmitted wirelessly via an artificial satellite for digital satellite broadcasting, or transmitted via a cable via a network such as a LAN (Local Area Network) or the Internet, and can be installed in the program memory 21.

In this case, when the program is upgraded, the upgraded program can be easily installed in the program memory 21.

Here, in this specification, processing steps for describing a program for causing the CPU 20 to perform various processes do not necessarily have to be processed in chronological order according to the order described in the flowchart, and may be performed in parallel. Alternatively, it also includes processing executed individually (for example, parallel processing or processing by an object).

Further, the program may be processed by one CPU or may be processed by a plurality of CPUs in a distributed manner.

Further, in the present embodiment, the continuous distribution HM
Although the speech recognition is performed in accordance with the M method, the speech recognition may be performed by, for example, matching between spectral waveforms in addition to the continuous distribution HMM method.

[0152]

According to the voice recognition apparatus, the voice recognition method, and the recording medium of the present invention, the voice recognition processing is controlled based on the state of the robot, so that it changes according to the state of the robot such as growth. The voice recognition processing can be realized, and as a result, it is possible to improve the entertainment property of the robot.

[Brief description of the drawings]

FIG. 1 is a perspective view illustrating an external configuration example of a robot according to an embodiment of the present invention.

FIG. 2 is a block diagram illustrating an internal configuration of the robot.

FIG. 3 is a block diagram illustrating a hardware configuration example of a control unit 11;

FIG. 4 is a block diagram illustrating a functional configuration example of the control unit 11, which is realized by the control unit 11 executing a program.

FIG. 5 is a diagram showing a growth model.

FIG. 6 is a diagram showing an action model.

FIG. 7 is a block diagram showing a configuration example of a voice recognition device as a portion for performing voice recognition in a sensor input processing unit 30;

FIG. 8 is a diagram showing a word dictionary.

FIG. 9 is a flowchart for explaining the operation of the speech recognition device in FIG. 7;

FIG. 10 is a diagram showing a dictionary growth score table.

FIG. 11 is a diagram showing a word dictionary for realizing “blur”.

[Explanation of symbols]

1 tail unit, 2 body unit, 3 head unit, 4A to 4D thigh units, 5A to 5D shin unit, 6A to 6D leg units 7 ₁ to 7 _N motor, 8 cameras, 9 microphone, 10 Pressure sensor, 11 control unit, 12 _{1 to} 12 _N rotary encoder, 20 CPU, 2
1 program memory, 22 RAM, 23 non-volatile memory, 24 I / F, 25 motor driver,
30 sensor input processing unit, 31 growth model unit, 3
2 Emotion / Instinct Model Department, 33 Action Decision Department, 33
A behavior model storage unit, 34 posture transition unit, 35 motor control unit, 41 feature parameter extraction unit, 42
Matching unit, 43 acoustic model storage unit, 44 dictionary storage unit, 45 grammar storage unit, 46 growth score setting unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 15/06 G10L 3/00 521F 15/10 521V 531G

Claims (11)

[Claims] 1. A voice recognition device built in a robot, comprising: voice recognition means for recognizing voice; and control means for controlling the voice recognition means based on a state of the robot. Voice recognition device. 2. The speech recognition apparatus according to claim 1, wherein the control means controls the speech recognition means based on a growth, an emotion, or an instinct state of the robot. 3. The speech recognition device according to claim 1, wherein the control unit changes the speech recognition accuracy of the speech recognition unit based on a state of the robot. 4. The speech recognition unit has a dictionary storage unit that stores a dictionary in which vocabulary to be subjected to speech recognition is registered, and the control unit is configured to control a vocabulary registered in the dictionary.
The voice recognition device according to claim 1, wherein the voice recognition unit is controlled so as to perform voice recognition with a weight corresponding to the state of the robot. 5. The speech recognition unit includes a dictionary storage unit that stores a plurality of dictionaries in which vocabularies to be subjected to speech recognition are separately registered. The control unit is configured to: The voice recognition device according to claim 1, wherein the voice recognition unit is controlled so as to perform voice recognition with a weight corresponding to the state of the robot. 6. The voice recognition unit includes a dictionary storage unit that stores a dictionary in which a vocabulary to be subjected to voice recognition is registered in association with another vocabulary. Controlling, based on the state, the voice recognition unit to output the other vocabulary associated with the vocabulary as a voice recognition result in the dictionary as a final voice recognition result. The voice recognition device according to claim 1. 7. A vocabulary to be subjected to speech recognition is registered in the dictionary in association with another vocabulary acoustically similar or semantically related to the vocabulary. The voice recognition device according to claim 6, wherein 8. The speech recognition means has a dictionary storage means for storing a dictionary in which vocabulary to be subjected to speech recognition is registered, and the control means registers in the dictionary based on a state of the robot. The speech recognition device according to claim 1, wherein the number of possible vocabularies is controlled. 9. The voice recognition device according to claim 1, wherein the robot takes a predetermined action based on a voice recognition result output by the voice recognition unit. 10. A voice recognition method for a voice recognition device built in a robot, comprising: a voice recognition step of recognizing voice; and a control step of controlling the voice recognition step based on a state of the robot. A speech recognition method characterized in that: 11. A recording medium storing a program to be executed by a computer for causing a robot to perform voice recognition, wherein the voice recognition step recognizes voice, and the voice based on a state of the robot. A recording medium characterized by recording a program comprising a control step for controlling a recognition step.